A sparse multifrontal solver using hierarchically semi-separable frontal matrices

Transcription

1 A sparse multifrontal solver using hierarchically semi-separable frontal matrices Pieter Ghysels Lawrence Berkeley National Laboratory Joint work with: Xiaoye S. Li (LBNL), Artem Napov (ULB), François-Henry Rouet (LBNL) 2014 CBMS-NSF Conference: Fast direct solvers for elliptic PDEs June 23 29, 2014, Dartmouth College

2 Introduction Consider solving Ax = b or M 1 A = M 1 b with a preconditioned iterative method (CG, GMRES, etc.) Fast convergence if good preconditioner M A is available 1. Cheap to construct, store, apply, parallelize 2. Good approximation of A Contradictory goals trade-off Standard design strategy Start with a direct factorization (like multifrontal LU) Add approximations to make it cheaper (cf. 1) while (hopefully/provably) affecting little (2) Approximation idea: use low-rank approximation

3 he multifrontal method [Duff & Reid 83] 5 Nested dissection reordering defines an elimination tree - SCOCH graph partitioner Bottom-up traversal of the e-tree At each frontal matrix, partial factorization and computation of a contribution block (Schur complement) Parent nodes sum the contribution blocks of their children: extend-add Elimination tree

4 Low-rank property In many applications, frontal matrices exhibit some low-rank blocks ( data sparsity ) Compression with SVD, Rank-Revealing QR,... - SVD: optimal but too expensive - RRQR: QR with column pivoting Exact compression or with a compression threshold ε Recursively, diagonal blocks have low-rank subblocks too - What to do with off-diagonal blocks?

5 Low-rank representations Most low-rank representations belong to the class of H-matrices [Bebendof, Börm, Hackbush, Grasedyck,... ]. Embedded in both dense and sparse solvers: U 1 V 1 U 2 U 3 = U1 0 U3 small 0 U2 Hierarchically Off-Diag. Low Rank (HODLR) [Ambikasaran, Cecka, Darve... ] U 6 B 36 V 3 Hierarchically Semiseparable (HSS) [Chandrasekaran, Dewilde, Gu, Li, Xia,... ] Nested basis. Block-low rank (BLR) [Amestoy et al.] Simple 2D partitioning. Recently in MUMPS. Also: H 2 (includes HSS and FMM), SSS... Choice: simple representations apply to broad classes of problems but provide less gains in memory/operations than specialized/complex ones.

6 HSS/HBS representation he structure is represented by a tree: 7 B 3,6 U 1 U 3, V 3 3 B 6,3 6 U 6, V 6 V 1 U 2 U 3 = U1 0 U3 small 0 U 2 B 1,2 B B 5,4 5 B 4,5 U 1, V 1 2,1 U 2, V 2 U 4, V 4 U 5, V 5 D 1 D 2 D 4 D 5 U 6 B 36 V 3 Number of leaves depends on the problem (geometry) and number of processors to be used. Building the HSS structure and all the usual operations (multiplying... ) consist of traversals of the tree.

7 Embedding low-rank techniques in a multifrontal solver 1. Choose which frontal matrices are compressed (size, level... ) 2. Low-rankness: weak interactions between distant variables = need suitable ordering/clustering of each frontal matrix Geometric setting (3D grid): 2D plane separator Need clusters with small diameters Hierarchical formats, merged clusters need small diameter too Split domain into squares and order with Morton ordering Algebraic: add some kind of halo to (complete) graph of separator variables and call a graph partitioner (MEIS) [Amestoy et al., Napov]

8 Embedding HSS kernels in a multifrontal solver More complicated HSS for frontal matrices: Fully structured: HSS on the whole frontal matrix. No dense matrix. Partial+: HSS on the whole frontal matrix. Dense frontal matrix. Partially structured: HSS on the F 11, F 12 and F 21 parts only. Dense frontal matrix, dense CB = F 22 F 21 F11 1 F 12 in stack. More memory F 11 F 21 F 12 F 22 Partially structured can do regular extend-add In partially structured, HSS compression of dense matrix After HSS compression, ULV factorization of F 11 block - Compared to classical LU in dense case Low rank Schur complement update

9 HSS compression via randomized sampling [Martinsson 11, Xia 13] HSS compression of a matrix A. Ingredients: R r and R c random matrices with d columns d = r + p with r estimated max rank; p = 10 in practice S r = AR r and S c = A R c samples of matrix A Can benefit from a fast matvec Interpolative Decomposition: A = A(:, J) X A is linear combination of selected columns of A wo sided ID: S c = S c (:, J c )X c and S r = S r (:, J r )X r, A = X c A(I c, I r )X r

10 HSS compression via randomized sampling 2 Algorithm (symmetric): from fine to coarse do Leaf node τ: 1. Sample: S loc = S(I τ, :) D R(I τ, :) 2. ID: S loc = U τ S loc (J τ, :) 3. Update: S τ = S loc (J τ, :) R τ = U τ R(I τ, :) I τ = I τ (J τ, :) Inner node τ with children [ ν 1, ν 2 : ] Sν1 A(I 1. Sample: S loc = ν1, I ν2 ) R ν2 S ν2 A(I ν2, I ν1 ) R ν1 2. ID: S loc = U τ S loc (J τ, :) 3. Update: S τ = S loc (J τ, :) R τ I τ = Uτ [R ν1 ; R ν2 ] = [I ν1 I ν2 ](J τ, :)

11 HSS compression via randomized sampling 3 If A A, do this for columns as well (simultaneously) [ ] I Bases have special structure: U τ = Π τ E τ Extract elements from frontal matrix: D τ = A(I τ, I τ ) and B ν1,ν 2 = A(I ν1, I ν2 ) Frontal matrix is combination of separator and HSS children Extracting element from HSS matrix requires traversing the HSS tree and multiplying basis matrices Limiting number of tree traversals is crucial for performance Benefits: Extend-add operation is simplified: only on random vectors Gains in complexity: O(r 2 N log N) iso O(rN 2 ) for non-randomized algorithm. log N due to extracting elements from HSS matrix

12 Randomized sampling extend-add Assembly in regular multifrontal: F p = A p CBc1 CBc2. Sample: S p = F p R p = (A p CBc1 CBc2 )R p = A p R p Y c1 Y c2 1D extend-add (only along rows); much simpler Y c1 and Y c2 samples of CB of children. R p = R c1 R c2 (+random rows for missing indices) Stages: Build random vectors from random vectors of children Build sample from samples of CB of children Multiply separator part of frontal matrix with random vectors: A p R p Compression of F p using S p and R p

13 HSS ULV factorization Exploit structure of U τ (from ID) to introduce zero s [ ] [ ] [ ] I Eτ I U τ = Π τ, Ω E τ = Π 0 τ I 0 τ Ω τ U τ = I D 1 U 1 B 2;1 V 2 U 2 B 1;2 V 1 D 2

14 HSS ULV factorization Exploit structure of U τ (from ID) to introduce zero s [ ] [ ] [ ] I Eτ I U τ = Π τ, Ω E τ = Π 0 τ I 0 τ Ω τ U τ = I [ ] [ ] [ ] Ω1 D 1 U 1B 1,2V2 W1 B = 1,2V Ω 2 U 2B 2,1V1 2 D 2 B 2,1V1 W 2 D 1 U 1 B 2;1 V 2 W 1;1 B 2;1 V 2 W 1;2 U 2 B 1;2 V 1 D 2 B 1;2 V 1 W 2;1 W 2;2

15 HSS ULV factorization Exploit structure of U τ (from ID) to introduce zero s [ ] [ ] [ ] I Eτ I U τ = Π τ, Ω E τ = Π 0 τ I 0 τ Ω τ U τ = I [ ] [ ] [ ] Ω1 D 1 U 1B 1,2V2 W1 B = 1,2V Ω 2 U 2B 2,1V1 2 D 2 B 2,1V1 W 2 ake (full) LQ decomposition [[ ] ] L 1 Lτ 0 Q W τ = τ [W 1;2 Q1 ] [B 1,2V2 Q 2 ] [ ] Q 1 W τ;2 L 2 Q 2 [B 2,1 V1 Q 1 ] [W 2;2Q2 ] D 1 U 1 B 2;1 V 2 W 1;1 B 2;1 V 2 L 1 B 2;1 V 2 Q 2 W 1;2 W 1;2,1 U 2 B 1;2 V 1 D 2 B 1;2 V 1 W 2;1 B 1;2 V 1 Q 1 L 2 W 2;2 W 2;2,1

16 HSS ULV factorization Exploit structure of U τ (from ID) to introduce zero s [ ] [ ] [ ] I Eτ I U τ = Π τ, Ω E τ = Π 0 τ I 0 τ Ω τ U τ = I [ ] [ ] [ ] Ω1 D 1 U 1B 1,2V2 W1 B = 1,2V Ω 2 U 2B 2,1V1 2 D 2 B 2,1V1 W 2 ake (full) LQ decomposition [[ ] ] L 1 Lτ 0 Q W τ = τ [W 1;2 Q1 ] [B 1,2V2 Q 2 ] [ ] Q 1 W τ;2 L 2 Q 2 [B 2,1 V1 Q 1 ] [W 2;2Q2 ] Rows for L τ can be eliminated, others are passed to parent At root node: LU solve of reduced D L 1 B 2;1 V 2 Q 2 W 1;2,1 B 1;2 V 1 Q 1 L 2 W 2;2,1

17 HSS ULV factorization Exploit structure of U τ (from ID) to introduce zero s [ ] [ ] [ ] I Eτ I U τ = Π τ, Ω E τ = Π 0 τ I 0 τ Ω τ U τ = I [ ] [ ] [ ] Ω1 D 1 U 1B 1,2V2 W1 B = 1,2V Ω 2 U 2B 2,1V1 2 D 2 B 2,1V1 W 2 ake (full) LQ decomposition [[ ] ] L 1 Lτ 0 Q W τ = τ [W 1;2 Q1 ] [B 1,2V2 Q 2 ] [ ] Q 1 W τ;2 L 2 Q 2 [B 2,1 V1 Q 1 ] [W 2;2Q2 ] Rows for L τ can be eliminated, others are passed to parent At root node: LU solve of reduced D ULV-like: Ω τ not orthonormal, forward/backward solve phases

18 Low rank Schur complement update Schur Complement update F 21 {}}{ F 22 F 21 F11 1 F 12 = F 22 U q B qk Vk F 1 11 F 12 {}}{ U k B kq Vq F 1 11 via ULV solve V k F 1 11 U k O(rN 2 ) D k is reduced HSS matrix O(r r) k q F 22 U q B qk (Ṽ k D 1 Ũ k )B kq V q F 11 F12 Ṽk D 1 Ũ k O(r 3 ) F 22 ΨΦ Ψ, Φ O(rN) U q and V q : traverse q subtree Cheap multiply with random vectors F 21 F 22

19 Numerical example general matrices GMRES(30) with right preconditioner Multifrontal with HSS vs ILUP from SuperLU b = (1, 1,... ), x 0 = (0, 0,... ) Stopping criterium: r k 2 / b HSS compression tolerance: 10 6 fill-ratio factor (s) its Matrix descr N rank HSS ILU HSS ILU HSS ILU add32 circuit 4, mchln85ks17 car tire 84, mhd500 plasma 250, poli large economics 15, stomach bio eng. 213, tdr190k accelerator 1, 100, torso3 bio eng. 259, utm5940 tokamak 5, wang4 device 26,

20 Numerical example tdr190k tdr190k: Maxwell equations in the frequency domain GMRES(30) convergence

21 Parallel implementation We have a serial code (StruMF [Napov ]) Some performance issues - Currently generates random vectors for all nodes in e-tree - How to estimate the rank? Currently guess and start over when too small Parallel implementation is a work in progress Distributed memory HSS compression of dense matrix - MPI, BLACS, PBLAS, BLAS, LAPACK Shared memory multifrontal code - OpenMP task parallelism for tree traversal for both elimination tree and HSS tree - Next step is parallel dense algebra

22 Parallel HSS compression MPI code opmost separator of a 3D problem, generated with exact multifrontal method k N 10,000 40,000 90,000 MPI processes / cores Nodes Levels olerance 1e-3 1e-3 1-e3 Non-randomized (s) Randomized (s) Dense LU ScaLAPACK (s) On 1024 cores achieved 5.3Flops/s very good flop balance: min / max = % communication overhead

23 ask based parallel tree traversal OpenMP Postorder (leaf to root) tree traversal: handle siblings in parallel, wait for children void traverse(node* p) { if (p->left) #pragma omp task traverse(p->left); if (p->right) #pragma omp task traverse(p->right); #pragma omp taskwait process(p->data); } Run-time system schedules tasks to cores (work-stealing) Root node will be handled sequentially: scaling bottleneck Nested trees: node of nested dissection tree contains HSS tree

24 time (s) time (s) ask based parallel tree traversal OpenMP HSS Compression of dense frontal matrix thread ID Multifrontal thread ID Blue: E-tree node, Red: HSS compression, Green: ULV-fact Extraction of elements from HSS matrix forms bottleneck Considering other runtime task schedulers Intel BB, StarPU, Quark he Quark scheduler from PLASMA could allow integration of PLASMA parallel (tiled) BLAS/LAPACK

25 Conclusions Developing an algebraic preconditioner for general nonsymmetric matrices (from PDEs) HSS is restricted format, large gains possible for certain applications, not for all Graph partitioning difficulties - Separator not always just a plane/line, not always a single piece - Bad for rank structure Some performance issues need to be addressed Separate distributed and shared memory codes - How to combine in a hybrid MPI+X code? Prepare for next generation NERSC supercomputer - Intel MIC based, > 60 cores

26 Conclusions Developing an algebraic preconditioner for general nonsymmetric matrices (from PDEs) HSS is restricted format, large gains possible for certain applications, not for all Graph partitioning difficulties - Separator not always just a plane/line, not always a single piece - Bad for rank structure Some performance issues need to be addressed Separate distributed and shared memory codes - How to combine in a hybrid MPI+X code? Prepare for next generation NERSC supercomputer - Intel MIC based, > 60 cores hank you! Questions?